UCLA

This thesis proposes to fully automate the shape and motion reconstruction of non-rigid objects from visual information using a unified deterministic/statistical deformable model. The model enhances the global control of a statistical deformable model with local control, based on assumptions of the material properties of the non-rigid object being modeled. A Histogram of Oriented Gradients (HoG) based object detector for a 3D volume is proposed to compute initial model estimates that are crucial for automation. This thesis also develops a unified variational method for 4D (3D+time) non-rigid shape reconstruction with anatomical and temporal smoothness constraints. The proposed unified model and method are combined in a fully automated Computer Vision and Machine Learning based framework for the clinically important application of segmenting the myocardium in cardiac cine Magnetic Resonance images.